Reading the Current: Tracking Subsea Cable Status through LLM-Assisted News Analysis

Abstract

Subsea Data Cables (SDCs) form the backbone of global digital infrastructure, handling over 99% of intercontinental data traffic. Still, they receive limited and dispersed media coverage, creating significant information gaps regarding their operational status. This paper aims to investigate the potential of Large Language Models (LLMs) in automating the extraction and analysis of SDC-related information from unstructured media sources. A comprehensive LLM-based information extraction module was developed and integrated into an existing SDC database. By systematically comparing different LLMs, including GPT-4o, Gemini 1.5 Flash, Claude 3.5 Sonnet, and Llama 3.1, this work identifies optimal model configurations for SDC news processing, considering accuracy, hallucination rate, and processing speed. The system implements Claude 3.5 Sonnet and GPT-4o as primary models, incorporating domain-specific prompt engineering and robust output validation mechanisms. A performance evaluation demonstrates substantial improvements over rule-based methods. While excelling at natural language processing tasks, the system revealed limitations in extracting more specific technical details such as construction costs and capacity measurements. The implementation provides a modular, adaptable framework for automated information extraction in specialised technical domains, including maritime ones. The results demonstrate that LLMs, when properly implemented with structured prompts and validation mechanisms, can significantly enhance the automated monitoring and analysis of SDC-related events.